Turning the PAGE - the potential of genome editing in breed-

This manuscript is a joint work of Henner Simianer¹, Torsten Pook¹ and Martin Schlather² and was presented at the World Congress on Genetics Applied to Live-stock Production 2018. For reasons of uniformity in this thesis, the conference/jour-nal style is not used in this section.

1: University of Goettingen, Animal Breeding and Genetics Group, Albrecht-Thaer-Weg 3, 37075 Goettingen, Germany

2: School of Business Informatics and Mathematics, University of Mannheim, A5, 68131 Mannheim, Germany

Author contributions by TP

TP performed the simulations for the study, contributed in the analysis and partic-ipated in revision of the manuscript.

Summary

In a recent study, Jenko et al. (2015) proposed to accelerate genetic progress by integrating a genome editing (GE) step in genomic breeding programs. This con-cept, called "promotion of alleles by genome editing" (PAGE) was implemented in a simulation study suggesting a substantial extra increase of genetic gain. As an example, editing in each generation the top 25 sires at the 20 quantitative trait nucleotides (QTN) with the highest eect was found to increase the genetic progress by 100% compared to genomic selection alone. We conducted a complex simulation study in which selection was on estimated GBLUP breeding values, the causal QTN were assumed unknown, and SNPs to be edited were identied by statistical means from the simulated data. We found the extra genetic progress due to PAGE to be be-tween 2 and 20 per cent, thus only a fraction of what was reported by Jenko et al.

(2015). The observed dierence is mainly attributed to the low power to detect true QTN. The best results were obtained with highly heritable traits, larger mapping populations and a limited number of true QTN, while performance was inferior both with low heritability traits and when QTN to be edited were identied by GWAS rather than by random regression BLUP. We argue that the true genetic architecture of complex traits will likely be much more complex than simulated here, which will further compromise the power of detecting true QTN and the predictability of the

eects of GE steps. These considerations together with the reported results indicate that overly optimistic expectations regarding the potential of PAGE should be taken with a pinch of salt.

Introduction

The concept of genome editing (GE) was rst successfully demonstrated in mammals in the 1970s, but only had a breakthrough with the introduction of the CRISPR-Cas9 system (Jinek et al., 2012). This approach combines simplicity, high accuracy, high eciency and limited o-target eects and has a considerable potential for mul-tiplexing. Based on this technological perspective Jenko et al. (2015) proposed to accelerate genetic progress by integrating a genome editing step in genomic breeding programs. The basic idea was to augment a genomic dairy cattle breeding scheme as originally suggested by Schaeer (2006) by changing a limited number of quantitative trait nucleotides (QTN) towards the most advantageous genotype in a given number of selected individuals via GE. This concept, called "promotion of alleles by genome editing" (PAGE), was demonstrated in a simulation study suggesting that, com-pared to a classical genomic selection (GS) scheme without a GE step, the expected response to selection with PAGE was between 1.08 and 4.12 times higher. As an example, editing in each generation the top 25 sires at the 20 QTN with the highest eect was found to double the genetic progress compared to GS alone.

While these results appear quite promising, a number of caveats must be made.

Most importantly, it was assumed that the true eects of segregating QTN were known and could be used to identify those QTN with the largest eects to be edited in the top sires. Also, selection was assumed to be based on known true breeding values.

The aim of our study was to assess the expected benet of PAGE with a somewhat more realistic scenario. For this, we set up a similar (but not completely identi-cal) simulation scheme and used genomic best linear unbiased prediction (GBLUP, (VanRaden, 2008)) for the selection step and estimated SNP eects from random regression BLUP (RRBLUP, (Meuwissen et al., 2001)) or alternatively GWAS re-sults to identify the top QTNs to be edited.

Simulation scheme

We started with a high density SNP data set from a real dairy cattle population which is assumed to reect the major characteristics of a breeding population under selection w.r.t. linkage disequilibrium, allele frequency spectra etc.. In the reference scenario we generated a base population of 500 male andN = 10⁰000female individ-uals. We selected 50'000 SNPs, of which 1000 SNPs (2%) were randomly assigned to be QTN with an additive eect drawn from a standard normal distribution. Phe-notypes of females were calculated as the sum of the genotype values across all QTN plus a random number sampled from a normal distribution such that the heritability

5.4 Potential of gene editing in breeding 145 had the desired value. We then used GBLUP to select the top 25 sires and 500 cows as potential bull sires (BS) and bull dams (BD), the remaining cows were used as cow dams (CD). In the rst ten generations, the 25 BS were randomly mated to the 500 BD to produce 500 male selection candidates for the next generation. The 25 BS were also mated to the 10'000 females (BD and CD) to produce 10'000 female selection candidates for the next generation.

Starting from generation 11, the 25 selected BS underwent a genome editing step in the PAGE approach: rst, eects were estimated for all SNPs using RRBLUP.

Next, each of the 25 BS was genome edited in that he was made homozygous for the favourable allele at those loci which had the highest estimated absolute SNP eects and for which he wasn't already carrying the most benecial genotype. These "edited bull sires" (eBS) were used in the same way as the BS in generation one to ten for ten more generations (11 20) (Figure 5.9)

Figure 5.9: One generation of the PAGE selection scheme (reference scenario).

Starting from this reference scenario, a number of parameters were varied, one at a time (reference value underlined):

ˆ the heritability: h²= 0.05, 0.3, 0.6

ˆ the number of true QTN: 1000, 500, 250

ˆ the number of SNPs: 50'000, 100'000, 200'000

ˆ the cow population size N: 10'000, 25'000

ˆ selection of target SNPs for editing: RRBLUP or GWAS

For each scenario we generated and analysed 100 replicates. The main criterion of interest was the change of genetic progress from generation 11 to 20 when using the PAGE approach compared to using genomic selection without GE in the same gen-erations. We also recorded the success rate of QTN identication as the proportion of edits that were made on true QTNs in each generation. The simulations were performed using the R-package RekomBre (Pook, unpublished)¹ on a server cluster with Intel E5-2670 (2X8 core 2.6 GHz) and AMD Opteron 6378 (4X16 2.4 GHz) processors.

Results and Discussion

Compared to selection based on GBLUP, we found that the PAGE approach led to an 11.6 per cent increase of the genetic progress in the reference scenario (Fig.

5.10 and 5.11), which is about one tenth of the improvement predicted by Jenko et al. (2015) for a similar scenario with 20 edits per top bull. Our implementation diers from the one described in Jenko et al. (2015) in many details, but we suspect two major causes for the observed discrepancy in the results: (i) while in Jenko et al. (2015) candidates are selected on their true breeding values, selection in our implementation is based on genomic breeding values, which are less accurate, but this aects selection based on GBLUP and PAGE both in a similar way; (ii) while loci to be edited are selected based on their true (but in reality unknown) eects in Jenko et al. (2015), selection is based on estimated allele eects from RRBLUP in our implementation. Especially the second discrepancy has a major eect, since in the rst generation of PAGE only 10 per cent of the edits are done on real QTN (see. Fig. 5.10), i.e. on average just 2 of the 20 loci edited in a bull were on target, while 18 edits were made on non-causal SNPs and thus had no consequence for the true genetic values of the edited bulls. This success rate even drops to 6.2 per cent in generation 20, since the detectable large eect QTNs tend to become xed in the rst generations of PAGE and the remaining polymorphic QTNs in later generations have smaller eects and are thus even more dicult to detect.

Varying some of the parameters in the reference scenario yielded the following obser-vations (Fig. 5.11): with a reduced number of QTN (500 or 250) the extra genetic progress due to PAGE increased to 14.5 and 15.5%, respectively, which can be ex-plained by the relatively larger eects and thus better detectability of true QTN.

However, with a smaller number of true QTN the success rate of QTN detection also tended to deteriorate faster. Increasing the number of SNPs had a slightly negative eect on the expected genetic progress. Increasing the population size to 25'000 cows led to a higher success rate in QTN detection and increased the extra genetic progress due to PAGE to 14.2%. Varying the heritability yielded the strongest ef-fects: while doubling h² to 0.6 led to an extra genetic progress of 20.7%, the latter was only 4.9% with h² = 0.05. An even more dramatic eect was observed when SNPs to be edited were identied based on GWAS results rather than RRBLUP

1RekomBre was used as the internal development title of MoBPS (Chapter 4)

5.4 Potential of gene editing in breeding 147

Figure 5.10: Genetic progress with PAGE (20 edits) vs. the pure GS variant (0 edits) and the proportion of edits on target in the GE generations (inner gure) in the reference scenario.

Figure 5.11: Extra genetic progress of PAGE vs. GS (left) and proportion of edits on target in generations 11 and 20 (right) for the dierent scenarios.

results. In this case, the proportion of edits of true QTN was between 3.2% and 3.7% and thus only marginally above the 2% expected when SNPs to be edited were selected completely at random, and consequently the extra genetic progress due to PAGE was only 2.5% in this scenario.

Conclusions

Across all scenarios studied, the increase of eciency due to PAGE didn't even get close to the rates suggested by the simulation study of Jenko et al. (2015), which can be mainly attributed to the low success rate of identifying true QTN for the editing step. One might expect that with increasing numbers of genotyped individuals the power to detect causal QTN will increase. However, the "true"

genetic model underlying our study as well as the one by Jenko et al. (2015) is heavily simplistic by assuming complete additivity. Genetic architecture of complex traits is expected to be far more complex in reality and will comprise epistatic interactions, complex nonlinear regulation processes and redundancies, genotype by environment and genotype by sex interactions etc. (Mackay, 2004). Causal variants will also be of more complex nature than being just single SNPs. All this will further reduce the power of nding true causal variants as targets for GE. Even if such a causal variant is detected and edited, the eect of this modication on the phenotype will be hardly predictable under such a complex genetic trait architecture. While GE presumably has some potential in breeding for monogenic traits, all these considerations together with the reported results indicate that overly optimistic expectations regarding the potential of PAGE in breeding for complex traits should be taken with a pinch of salt.